An optimised YOLOv4 deep learning model for efficient malarial cell detection in thin blood smear images

Area of Science:

• Bioinformatics
• Medical Imaging
• Deep Learning

Background:

• Malaria diagnosis is critical for global health, but manual methods are slow.
• Automated tools are needed to expedite malaria detection and treatment.
• This study focuses on improving deep learning models for malaria diagnosis.

Purpose of the Study:

• To develop a more lightweight and accurate YOLOv4 model for malaria diagnosis.
• To enhance the speed and precision of identifying infected red blood cells.
• To reduce the computational complexity of deep learning models in medical diagnostics.

Main Methods:

• Modified the YOLOv4 model using direct layer pruning and backbone replacement.
• Replaced the CSP-DarkNet53 backbone with a shallower ResNet50 network for feature extraction.
• Analyzed and compared performance metrics of the modified models against the original YOLOv4.

Main Results:

• The modified YOLOv4-RC3_4 model achieved a mean accuracy precision (mAP) of 90.70%, an improvement of over 9% compared to the original model.
• Achieved approximately 22% reduction in computations and 23 MB in model size.
• Demonstrated a 9.27% increase in detecting infected cells after pruning redundant layers.

Conclusions:

• The YOLOv4 model, with pruned residual blocks, effectively detects infected red blood cells.
• This optimization reduces model size and computational cost without sacrificing accuracy.
• The developed method offers a potential revolution in malaria diagnosis and bioinformatics solutions.